A Multi-source DEM Point-Surface Fusion Model Based on Deep Belief Network

  Objectives  Although there have been a variety of global digital elevation model (DEM) products available for geospatial applications, the data quality of DEMs are inevitably affected by the observing techniques and processing methods. Multi-source data fusion by integrating the complementary information among the datasets is an effective way to improve DEM data quality. Based on the facts, this paper proposed a point-surface fusion model based on deep belief network (DBN).

  Methods  The terrain slope, land cover, and the spatial coordinates were integrated into the DBN model, thus constructed the statistical relation between the DEM elevations and the high-quality ICESat (ice, cloud, and land elevation) GLAS (geoscience laser altimeter system) data. The DBN-based fusion model was capable to learn the spatial structure and landscape associations of the DEM errors, and thus the multi-source raster DEMs and the high-quality LiDAR altimetry can be effectively fused. For the data tiles with voids, the triangulated irregular network (TIN) delta surface method was employed to merge the fusion results for the void and non-void areas, respectively.

  Results  In the experiments, the effectiveness of the proposed method is tested on the fusion of ASTER (advanced spaceborne thermal emission and reflection radiometer) GDEM (global digital elevation model) v2, ALOS AW3D30 (advanced land observing satellite world 3D-30 m) and SRTM1 (shuttle radar topography mission 1-arc second) data, which are the most popular DEM products. Six data tiles with different terrain conditions are tested, and the accuracy of the original data and the fusion results are evaluated using the ICESat GLAS points with cross-validation. Quantitative results show that the fused data have higher accuracy compared with the original datasets and the pair-wise fusion results, both in terms of the void areas and the non-void areas.

  Conclusions  The results show that the proposed method is able to improve the accuracy of the DEM data with fusion of multi-source altimetry datasets. Moreover, it is effective in dealing with the data voids, noise and abnormal values simultaneously.